[PATCH 0/3] rpdfs-progs: ARM64 support and documentation

[PATCH 0/3] rpdfs-progs: ARM64 support and documentation

[Valerie Aurora]

A couple of hacks to support ARM64 and some documentation updates.

The ARM64 patches should NOT be merged into main - I just wanted to
share them here in case anyone else wanted to use them.

Valerie Aurora (3):
  Initial support for ARM64, incompatible with x86_64
  Workaround for short read from io_uring on aarch64
  Update documentation

 Makefile            |  9 +++--
 docs/development.md | 16 +++------
 utask/net.c         | 14 ++++++--
 utask/utask.c       | 17 ++++++++-
 utask/utask_asm.S   | 84 +++++++++++++++++++++++++++++----------------
 utask/utask_defs.h  |  2 +-
 6 files changed, 94 insertions(+), 48 deletions(-)

-- 
2.49.0

[PATCH 1/3] Initial support for ARM64, incompatible with x86_64

[Valerie Aurora]

A quick hack to support task switching on ARM64 without any attempt to
make it portable or compatible with x86_64. Will come back to it later
if there is interest.
---
 Makefile           |  9 +++--
 utask/utask.c      | 17 +++++++++-
 utask/utask_asm.S  | 84 ++++++++++++++++++++++++++++++----------------
 utask/utask_defs.h |  2 +-
 4 files changed, 79 insertions(+), 33 deletions(-)

diff --git a/Makefile b/Makefile
index 8253885..6511202 100644
--- a/Makefile
+++ b/Makefile
@@ -3,7 +3,9 @@
 # make every target depend on the makefile
 .EXTRA_PREREQS:= $(abspath $(lastword $(MAKEFILE_LIST)))
 
-CFLAGS := -I. -O2 -ggdb -Wall -Werror -D_FILE_OFFSET_BITS=64 -msse4.2 -fno-strict-aliasing -fno-omit-frame-pointer
+#CFLAGS := -I. -O2 -ggdb -Wall -Werror -D_FILE_OFFSET_BITS=64 -msse4.2 -fno-strict-aliasing -fno-omit-frame-pointer
+
+CFLAGS := -I. -I/usr/lib64/ -O2 -ggdb -Wall -Werror -D_FILE_OFFSET_BITS=64 -fno-strict-aliasing -fno-omit-frame-pointer
 
 # function entrance/exit profiling for uftrace
 CFLAGS += -pg
@@ -16,6 +18,9 @@ LDFLAGS := -Wl,--gc-sections -lxxhash -luring -luuid -lsystemd
 # provide dynamic symbol tables for backtrace()
 LDFLAGS += -rdynamic
 
+# aarch64 needs this to find libraries
+LDFLAGS += -L/usr/lib64/
+
 # build all c files in source directories
 DIR := cli devd shared shared/lk utask
 SRC := $(foreach d,$(DIR),$(wildcard $(d)/*.c))
@@ -73,7 +78,7 @@ $(OBJ_S): %.o: %.S utask/utask_defs.h utask/utask_gen_defs.h
 utask/utask.s: utask/utask.c
 	gcc -I. -S -o ./utask/utask.s  ./utask/utask.c
 
-# extract defines from complied asm so we can include it from asm source
+# extract defines from compiled asm so we can include it from asm source
 utask/utask_gen_defs.h: utask/utask.s
 	grep '#define.*\<UTASK_ASM' utask/utask.s > utask/utask_gen_defs.h
 
diff --git a/utask/utask.c b/utask/utask.c
index 7006f5c..3575b67 100644
--- a/utask/utask.c
+++ b/utask/utask.c
@@ -97,7 +97,7 @@ static struct utask *get_current_utask(void)
 	struct utask *tsk;
 	unsigned long sp;
 
-	asm("movq %%rsp, %0 		\n\t"
+	asm("mov %0, sp 		\n\t"
 	    : "=rm" (sp) : : );
 
 	tsk = (void *)(round_up(sp, UTASK_STACK_SIZE) - sizeof(struct utask));
@@ -405,6 +405,7 @@ int utask_create_name_nowake(char *name, utask_fn_t fn, void *data, struct utask
 	struct utask *tsk = NULL;
 	unsigned long after;
 	void *stack;
+	void *fp;
 	int ret;
 
 	ret = posix_memalign(&stack, UTASK_STACK_SIZE, UTASK_STACK_SIZE);
@@ -439,15 +440,29 @@ int utask_create_name_nowake(char *name, utask_fn_t fn, void *data, struct utask
 	if (!IS_ALIGNED(after, 16))
 		tsk->sp -= 16 - (after & 15);
 
+	/* for alignment (I think?) */
+	push_stack(tsk, NULL);
+
 	/* final null address to terminate backtracing unwinders */
 	push_stack(tsk, NULL);
+
 	/* first switch_to jumps to _entry */
 	push_stack(tsk, utask_entry);
 
+	/* aarch64 wants the return address and the frame pointer on the stack */
+	fp = tsk->sp;
+	push_stack(tsk, fp);
+
 	/* initial zeroed saved registers */
 	tsk->sp -= UTASK_SAVED_BYTES;
 	memset(tsk->sp, 0, UTASK_SAVED_BYTES);
 
+	/* aarch64 - write the frame pointer into the second to last register */
+	*(void **)(tsk->sp) = fp;
+
+	/* aarch64 - write the return address into the last register (link register) */
+	*(void **)(tsk->sp + sizeof(void *)) = utask_entry;
+
 	ret = 0;
 out:
 	if (ret < 0)
diff --git a/utask/utask_asm.S b/utask/utask_asm.S
index a693520..c69a090 100644
--- a/utask/utask_asm.S
+++ b/utask/utask_asm.S
@@ -14,46 +14,48 @@
 # and epilogue.  So we save and restore them as we switch to and from
 # stacks.
 #
-# ``Registers %rbp, %rbx and %r12 through %r15 “belong” to the calling
+# ``Registers %rbp, %rbx and %r12 through %r15 "belong" to the calling
 # function and the called function is required to preserve their
 # values.'' -- x86_64 ABI.
 #
 .macro push_callee_saved
-	push %rbx
-	push %r12
-	push %r13
-	push %r14
-	push %r15
-	push %rbp
+	stp	x19, x20, [sp, #-16]!
+	stp	x21, x22, [sp, #-16]!
+	stp	x23, x24, [sp, #-16]!
+	stp	x25, x26, [sp, #-16]!
+	stp	x27, x28, [sp, #-16]!
+	stp	x29, x30, [sp, #-16]!
 .endm
 
 .macro pop_callee_saved
-	pop %rbp
-	pop %r15
-	pop %r14
-	pop %r13
-	pop %r12
-	pop %rbx
+	ldp	x29, x30, [sp], #16
+	ldp	x27, x28, [sp], #16
+	ldp	x25, x26, [sp], #16
+	ldp	x23, x24, [sp], #16
+	ldp	x21, x22, [sp], #16
+	ldp	x19, x20, [sp], #16
 .endm
 
 #
-# set rbx to the address of the utask struct found at the base of the
+# set x9 to the address of the utask struct found at the base of the
 # stack.
 #
-.macro load_utask_rbx
-	movq %rsp, %rbx
-	andq $UTASK_STACK_BASE_MASK, %rbx
-	addq $(UTASK_STACK_SIZE - UTASK_ASM_SIZEOF_UTASK), %rbx
+.macro load_utask_x9
+	mov	x9, sp
+	mov	x10, #UTASK_STACK_BASE_MASK
+	and	x9, x9, x10
+	mov	x10, #(UTASK_STACK_SIZE - UTASK_ASM_SIZEOF_UTASK)
+	add	x9, x9, x10
 .endm
 
 #
-# int utask_switch_to(utask *utask /* rdi */);
+# int utask_switch_to(utask *utask /* x0 */);
 #
 # Resume execution of a utask that was previously stopped by switching
 # from the scheduler's stack to the task's.
 #
 # Save the calling scheduler's stack pointer, restore the utask's stack
-# pointer, and return into the address that the utask pushed as they
+# pointer, and return into the address that the utask saved as they
 # last called _switch_from.
 #
 # The ret here is returning to the caller of _switch_from, which is void,
@@ -63,8 +65,18 @@
 .type utask_switch_to, @function
 utask_switch_to:
 	push_callee_saved
-	movq %rsp, UTASK_ASM_OFFSETOF_SCHED_SP(%rdi)
-	movq UTASK_ASM_OFFSETOF_SP(%rdi), %rsp
+	// store scheduler/caller sp into utask struct at x0
+	mov x9, x0
+	add x9, x9, UTASK_ASM_OFFSETOF_SCHED_SP
+	mov x10, sp
+	str x10, [x9]
+
+	// load utask sp from utask struct into sp
+	mov x9, x0
+	add x9, x9, UTASK_ASM_OFFSETOF_SP
+	ldr x10, [x9]
+	mov sp, x10
+
 	pop_callee_saved
 	ret
 
@@ -83,11 +95,22 @@ utask_switch_to:
 .type utask_switch_from, @function
 utask_switch_from:
 	push_callee_saved
-	load_utask_rbx
-	movq %rsp, UTASK_ASM_OFFSETOF_SP(%rbx)
-	movq UTASK_ASM_OFFSETOF_SCHED_SP(%rbx), %rsp
+	load_utask_x9 // load the calling utask's utask struct into x9
+
+	// store the utask sp into utask struct
+	mov x10, x9
+	add x10, x10, UTASK_ASM_OFFSETOF_SP // add the utask sp offset
+	mov x11, sp
+	str x11, [x10] // store sp in utask sp
+
+	// load the scheduler sp into the sp of the utask struct in x9
+	mov x10, x9
+	add x10, x10, UTASK_ASM_OFFSETOF_SCHED_SP // add the sched sp offset
+	ldr x11, [x10]
+	mov sp, x11
+
 	pop_callee_saved
-	movq $UTASK_RET_SCHEDULED, %rax
+	mov x0, UTASK_RET_SCHEDULED
 	ret
 
 #
@@ -102,8 +125,11 @@ utask_switch_from:
 .globl utask_finish
 .type utask_finish, @function
 utask_finish:
-	load_utask_rbx
-	movq UTASK_ASM_OFFSETOF_SCHED_SP(%rbx), %rsp
+	// load scheduler sp out of utask struct into sp
+	load_utask_x9
+	add x9, x9, UTASK_ASM_OFFSETOF_SCHED_SP // add the sched sp offset
+	ldr x10, [x9]
+	mov sp, x10
 	pop_callee_saved
-	mov $UTASK_RET_FINISHED, %rax
+	mov x0, UTASK_RET_FINISHED // return UTASK_RET_FINISHED
 	ret
diff --git a/utask/utask_defs.h b/utask/utask_defs.h
index 5f186c3..f5cd62d 100644
--- a/utask/utask_defs.h
+++ b/utask/utask_defs.h
@@ -11,7 +11,7 @@
 #define UTASK_STACK_SIZE	(8 * 1024)
 #define UTASK_STACK_BASE_MASK	(~(UTASK_STACK_SIZE - 1))
 
-#define UTASK_SAVED_BYTES	(6 * 8)
+#define UTASK_SAVED_BYTES	(12 * 8)
 
 #define UTASK_RET_SCHEDULED	0
 #define UTASK_RET_FINISHED	1
-- 
2.49.0

Re: [PATCH 1/3] Initial support for ARM64, incompatible with x86_64

[Zach Brown]

On Tue, Feb 24, 2026 at 05:33:24PM +0100, Valerie Aurora wrote:
> A quick hack to support task switching on ARM64 without any attempt to
> make it portable or compatible with x86_64. Will come back to it later
> if there is interest.

Fun!  I know of one ponential user that might want to experiment with
rpdfs once its up and running and who would be using ARM.  Maybe we'll
be able to explore this with them when the time comes.

- z

[PATCH 2/3] Workaround for short read from io_uring on aarch64

[Valerie Aurora]

For some reason, io_uring returns a short read even with MSG_WAITALL -
but on aarch64 and not x86_64. Work around that with an outer loop
---
 utask/net.c | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/utask/net.c b/utask/net.c
index 7f21e4c..35919fd 100644
--- a/utask/net.c
+++ b/utask/net.c
@@ -265,11 +265,19 @@ static ssize_t recv_waiter(int fd, void *buf, size_t size, int flags)
 {
 	struct utask_cqe_waiter waiter;
 	struct io_uring_sqe *sqe;
+	size_t len = 0;
 	int ret;
 
-	sqe = utask_get_sqe_waiter(&waiter);
-	io_uring_prep_recv(sqe, fd, buf, size, flags | MSG_WAITALL);
-	ret = utask_wait_event_task(waiter.completed);
+	/* TODO: Figure out why io_uring returns a short read on aarch64 */
+	while (len < size) {
+		sqe = utask_get_sqe_waiter(&waiter);
+		io_uring_prep_recv(sqe, fd, buf + len, size - len, flags | MSG_WAITALL);
+		ret = utask_wait_event_task(waiter.completed);
+
+		if (waiter.res < 0)
+			break;
+		len += waiter.res;
+	}
 
 	/* == 0 when fd is shutdown/disconnected */
 	BUG_ON(ret == 0 && waiter.res > 0 && waiter.res != size);
-- 
2.49.0

Re: [PATCH 2/3] Workaround for short read from io_uring on aarch64

[Zach Brown]

> +	/* TODO: Figure out why io_uring returns a short read on aarch64 */

That is certainly peculiar.  I wouldn't be surprised if we're not using
io_uring net correctly.  What's there is very minimal.

- z

[PATCH 3/3] Update documentation

[Valerie Aurora]

---
 docs/development.md | 16 ++++------------
 1 file changed, 4 insertions(+), 12 deletions(-)

diff --git a/docs/development.md b/docs/development.md
index df606b4..02974dd 100644
--- a/docs/development.md
+++ b/docs/development.md
@@ -25,9 +25,7 @@ Example invocation:
 In another terminal:
 
 ```
-	$ ./cli/rpdfs-cli debugfs -d 127.0.0.1:8081 -t /tmp/trace_debugfs
-	<1> $ mkfs
-	<1> $ stat
+	# mount -t rpdfs -o mkfs 127.0.0.1:8081 /mnt
 ```
 
 ## Code Layout
@@ -43,18 +41,12 @@ This is all the code that's shared by each utility.  It's not a proper
 library in that it doesn't need to remain API compatible with external
 builds over time.
 
-There's two kinds of shared code.  There's code that can run in either
-userspace or the kernel (block.c) and shared code that only runs in
-userspace (options.c).  It'd probably be worth making this distinction
-more apparent.
-
 **shared/lk/**
 
-This is for userspace implementations of kernel interfaces.  This both
+This is for userspace implementations of kernel interfaces. This both
 lets us use reasonably stand-alone kernel interfaces (list.h) in
-userspace as well as share rpdfs code with the kernel module by
-providing implementations of more complicated runtime services (RCU hash
-tables, work queues).
+userspace by providing implementations of more complicated runtime
+services (RCU hash tables, work queues).
 
 **shared/format-{block,msg,trace}.h**
 
-- 
2.49.0

Re: [PATCH 3/3] Update documentation

[Zach Brown]

On Tue, Feb 24, 2026 at 05:33:26PM +0100, Valerie Aurora wrote:
> ---
>  docs/development.md | 16 ++++------------
>  1 file changed, 4 insertions(+), 12 deletions(-)

Hmm, can you add a Signed-off-by: so I can merge this?  (And maybe clean
up that edit of the last paragraph.. it's a little fractured.  Don't
sweat it though, I was going to take another editing pass on top of
these cleanups either way.)

- z